Descriptor: A High-Fidelity 3-D Dataset for Fine-Grained Monocular Depth Estimation (HFD-3D)

This article presents the High-Fidelity 3D Dataset for Fine-Grained Monocular Depth Estimation (HFD-3D), a new resource for developing and evaluating depth estimation models. The primary objective of HFD-3D is to provide a benchmark with verified sub-millimeter ground-truth fidelity, addressing a key limitation in existing datasets, which often provide depth errors on the order of centimeters. The dataset comprises endoscopic RGB images paired with dense, metric-accurate ground truth depth maps. The scenes feature a diverse range of objects designed to test model generalization, including basic geometric phantoms, complex surgical task simulators, soft fabrics, porous foams, and ex-vivo tissues. Ground truth depth was acquired using a high-precision 3D scanner and a multi-stage processing pipeline to ensure high fidelity and precise alignment between the 2D images and 3D data. HFD-3D is designed to thoroughly test the ability of algorithms to reconstruct fine-grained details and handle surfaces with complex light interaction, such as translucency and specularity. To advance monocular 3D reconstruction for high-precision applications, this work offers a standardized benchmark. We establish comprehensive baseline performance metrics using a fine-tuned DepthAnythingV2 model alongside zero-shot evaluations of state-of-the-art architectures. Furthermore, we provide pipeline-component ablations and demonstrate cross-dataset generalization on the C3VD benchmark. The complete open-source pipeline is also provided, ensuring reproducibility and allowing the community to customize or extend the dataset for future research.